OpenAI’s AI Models Have a Goblin Problem, and They’re Finally Talking About It

OpenAI finally fessed up about the goblins. You know, the whole thing where Wired caught their coding model being told to “never talk about goblins, gremlins, raccoons, trolls, ogres, pigeons, or other animals or creatures.” That raised some eyebrows, and not just because it sounds like a rejected Monty Python sketch.

So OpenAI published a blog post calling it a “strange habit” their models picked up during training. They started noticing it with GPT-5.1, specifically when users selected the “Nerdy” personality option. And it got worse with each subsequent model refresh. The models would just start dropping metaphors about goblins and other fantasy creatures into responses, unprompted. It’s like they couldn’t help themselves.

I’ve seen this kind of thing before with large language models. They latch onto weird patterns in training data that humans would never notice or think to filter out. But goblins? That’s a new one for me. It’s almost charming in a way, but also a reminder that these systems are still fundamentally unpredictable.

The fix? OpenAI says they’re tweaking the training pipeline to reduce these metaphorical outbursts. They’re not banning the creatures outright—that would be ironic given the original instruction—but they’re adjusting the model’s priors so it doesn’t default to goblin talk when it’s trying to sound clever.

Honestly, I’m a bit disappointed. Gremlins in the code were the least of our problems, and now we’ve lost a source of unintentional entertainment. But I get it: consistency matters more than quirky metaphors when you’re shipping a product. Still, I’ll miss the odd goblin reference. It gave the models some personality, even if it was a bug.